ASVspoof5
收藏资源简介:
ASVspoof 5(Track 1,评估集)是ASVspoof 5挑战赛中用于语音反欺骗和合成/深度伪造语音检测任务的基准数据集,专门包含Track 1的完整评估分区。该数据集旨在支持语音防伪研究,任务为二分类:区分真实人类语音(bonafide)与合成或转换的欺骗语音(spoof)。数据集规模为680,774个语音片段(trials),其中包含138,688个真实样本和542,086个欺骗样本。数据以表格形式组织,包含四个核心字段:path(唯一文件名,格式为<utterance_id>.flac)、audio(16 kHz单声道FLAC格式的原始音频数据,未重新编码)、label(分类标签,bonafide对应0,spoof对应1)以及notes(JSON格式的元数据字符串,包含话语ID、说话人ID、性别、编解码器、编解码器ID、源ID、攻击条件和攻击ID等信息)。该数据集基于原始ASVspoof挑战赛数据,按照Open Data Commons Attribution License (ODC-By) v1.0许可重新分发,适用于语音深度伪造检测、音频分类等任务的模型训练与评估。
ASVspoof 5 (Track 1, Evaluation Set) is a benchmark dataset from the ASVspoof 5 challenge for speech anti-spoofing and synthetic/deepfake speech detection tasks, specifically containing the complete evaluation partition for Track 1. The dataset is designed to support speech anti-counterfeiting research, with a binary classification task: distinguishing between genuine human speech (bonafide) and synthetic or converted spoof speech. The dataset comprises 680,774 speech trials, including 138,688 bonafide samples and 542,086 spoof samples. Data is organized in a tabular format with four core fields: path (unique filename in the format <utterance_id>.flac), audio (raw audio data in 16 kHz mono FLAC format, not re-encoded), label (classification label, where bonafide corresponds to 0 and spoof to 1), and notes (a JSON-formatted metadata string containing utterance ID, speaker ID, gender, codec, codec ID, source ID, attack condition, and attack ID). The dataset is based on the original ASVspoof challenge data and is redistributed under the Open Data Commons Attribution License (ODC-By) v1.0, suitable for model training and evaluation in tasks such as speech deepfake detection and audio classification.
数据集概述
ASVspoof 5 (Track 1, Eval) 是一个面向音频深度伪造检测的基准测试数据集,专门用于语音反欺骗和合成/深度伪造语音检测任务。
- 任务类型:音频二分类(真实语音 vs. 欺骗语音)
- 标签:
bonafide(真实人声,标签0) /spoof(合成/转换语音,标签1) - 数据规模:总计 680,774 条样本,其中真实语音 138,688 条,欺骗语音 542,086 条
- 数据格式:16 kHz 单声道 FLAC 音频,嵌入在数据集中(位精确解码验证通过)
- 许可协议:Open Data Commons Attribution License (ODC-By) v1.0
数据模式(Schema)
| 列名 | 类型 | 说明 |
|---|---|---|
path |
string |
音频文件名(<utterance_id>.flac) |
audio |
Audio(16000) |
16 kHz 单声道 FLAC 音频 |
label |
ClassLabel |
"bonafide" (0) / "spoof" (1) |
notes |
string |
JSON 格式元数据,包含:utterance_id、speaker_id、gender、codec、codec_id、source_id、attack_condition、attack_id |
快速使用示例
python from datasets import load_dataset
ds = load_dataset("SpeechAntiSpoofingBenchmarks/ASVspoof5", split="test") print(ds[0])
关键统计信息
| 指标 | 数值 |
|---|---|
| 总样本数 | 680,774 |
| 真实语音(Bonafide) | 138,688 |
| 欺骗语音(Spoof) | 542,086 |
数据来源与引用
- 原始挑战赛:https://www.asvspoof.org/
- 评估协议文件:
ASVspoof5.eval.track_1.tsv - 相关论文:Wang et al., "ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale", ASVspoof Workshop 2024.
- 维护者联系方式:k.n.borodin@mtuci.ru




